Chronic wounds—defined clinically as persistent skin injuries failing to progress through the normal, orderly stages of healing within four weeks—represent an escalating global health crisis. These stubborn lesions are frequently intertwined with complex underlying systemic pathologies, including diabetes mellitus, vascular insufficiencies, pressure necrosis, and major trauma. Beyond the severe physical pain and emotional distress imposed on patients, chronic wounds exact a staggering economic toll on healthcare infrastructure worldwide.
In a recent milestone retrospective cohort study conducted at Chongqing University Qianjiang Hospital, researchers set out to untangle the intricate physiological, social, and psychological determinants governing chronic wound prognosis. By analyzing data from 232 patients treated between January 2023 and June 2024, the investigative team deployed a dual-analytical framework. They integrated traditional statistical tools—such as binary logistic regression—with modern machine learning architectures, specifically Multilayer Perceptron (MLP) neural networks.
The study’s core findings underscore that patient nutritional status, as measured by serum albumin levels, and comprehensive, multimodal clinical interventions are vital prognostic pillars. Furthermore, the research demonstrates that advanced computational models can match the diagnostic accuracy of traditional statistical methods while unlocking complex, non-linear relationships among physical wound parameters.
Detailed Chronology & Methodology
To bridge existing knowledge gaps regarding the non-linear biological processes of wound healing, the research team structured a rigorous retrospective observational study.
Patient Cohort and Inclusion Parameters
The study evaluated medical and demographic records from 232 patients who presented with wounds persisting for a minimum of four weeks at Chongqing University Qianjiang Hospital. To maintain data integrity, individuals with acute wounds under four weeks of duration or those lacking core clinical records that could not be supplemented via telephone follow-up were excluded. The institutional review board officially approved the protocol (No: CQSQJZXYY-2025162), waiving formal written consent due to the retrospective, de-identified nature of the clinical data.
Categorization of Wound Prognosis
At the termination of a minimum 12-week follow-up period, patient outcomes were stratified into two distinct categories:
- Good Prognosis: Defined as complete epithelial closure or a clinically documented reduction in wound surface area of 50% or greater without recurrence.
- Poor Prognosis: Defined as an absence of healing, less than a 50% area reduction, clinical deterioration, or active recurrence.
Out of the 232 total participants, 122 patients (52.6%) achieved a good prognosis, while 110 patients (47.4%) experienced a poor prognosis.
Analytical Pipeline: From Univariate to Machine Learning
The analytical approach unfolded in sequential phases:
- Univariate Analysis: Continuous variables were evaluated using the Mann-Whitney U test, categorical variables via chi-square or Fisher’s exact tests, and ordered variables through the Cochran-Armitage trend test. This screening identified seven statistically significant parameters ($P < 0.05$): number of concurrent wounds, initial wound size, pain score, intervention modalities, non-steroidal anti-inflammatory drug (NSAID) use, hemoglobin, and serum albumin.
- Multivariable Binary Logistic Regression: All seven significant univariate factors were entered simultaneously into a regression model. The model controlled for confounding variables while evaluating adjusted odds ratios (ORs).
- Multilayer Perceptron (MLP) Neural Network: To capture non-linear patterns, an MLP model was constructed featuring an input layer of seven neurons, two hidden layers (eight and four neurons utilizing ReLU activation), and a single sigmoid output neuron. The dataset was split into a training set ($n = 160$, 69.0%) and a held-out test set ($n = 72$, 31.0%), validated via 10-fold cross-validation.
Supporting Context, Metrics, and Findings
Nutritional Biomarkers and Systemic Health
The investigation revealed that serum albumin is a powerful independent predictor of healing success. In the multivariable logistic regression model, every 1 g/L increase in serum albumin improved the likelihood of a good prognosis by approximately 10% (adjusted OR = 1.101, 95% CI: 1.024–1.183, $P = 0.010$). Furthermore, in the MLP neural network, albumin emerged as the second most influential feature, boasting a normalized importance score of 69.3%.
While hemoglobin levels initially demonstrated significance in univariate screening ($P = 0.005$), they lost independent predictive power after multivariable adjustment ($P = 0.504$). Multicollinearity diagnostics confirmed this was not due to statistical distortion (Variance Inflation Factors for both markers remained below 1.7), but rather because albumin serves as a broader, more comprehensive indicator of underlying systemic inflammation, protein synthesis capacity, and nutritional competence.
The Impact of Multimodal Interventions
Treatment strategies analyzed across the cohort included routine dressing changes, surgical debridement, targeted infection control, management of underlying comorbidities, laser therapy, and structured nutritional support.
The data revealed a striking trend: patients receiving two or three distinct intervention modalities experienced significantly higher rates of favorable healing compared to those receiving no structured intervention (adjusted OR = 4.775 and 6.360, respectively). However, the researchers emphasize that this correlation reflects prognostic patterns rather than definitive causation. Patients presenting with more complex clinical profiles frequently warranted intensive, multi-pronged care strategies.
Wound Metrics and Neural Network Insights
Wound size played a pivotal role in shaping patient trajectories. In univariate analysis, patients with a poor prognosis presented with significantly larger median wound sizes (9 $textcm^2$) compared to those with a good prognosis (6 $textcm^2$, $P = 0.043$).
Within the MLP architecture, initial wound size claimed the top spot with 100% normalized importance. This highlights a distinct advantage of machine learning over standard linear models: neural networks are uniquely equipped to identify threshold effects and non-linear interactions where oversized wounds overwhelm local tissue regeneration capacities.
Predictive Accuracy and Model Comparison
Both analytical models demonstrated robust discriminative capability:
- Logistic Regression: Achieved an Area Under the Curve (AUC) of 0.773 (95% CI: 0.713–0.833) with an overall classification accuracy of 69.8%.
- MLP Neural Network: Achieved an AUC of 0.771 (95% CI: 0.710–0.832) on the held-out test set, with overall test accuracy hitting 68.1%.
A DeLong test comparing the two AUCs yielded $P = 0.907$, confirming that the machine learning model achieved diagnostic performance equivalent to traditional statistical regression while offering superior insight into complex, multi-variable interactions.
Official Statements and Institutional Perspectives
The research team, led by corresponding authors Yuping Ran and Meifeng Xu alongside lead investigator Fang Zhang, emphasized the clinical implications of their findings while maintaining rigorous scientific caution.
"Our findings underscore the fundamental importance of patient nutritional status and structured, comprehensive wound management protocols," the authors noted in their official study conclusions. "However, given the single-center, retrospective observational design, all documented associations must be interpreted as prognostic patterns rather than direct causal pathways."
The institutional review board at Chongqing University Qianjiang Hospital affirmed that the study adheres strictly to ethical standards for retrospective medical research. The board highlighted that anonymizing historical clinical datasets protects patient confidentiality while allowing vital epidemiological insights to emerge from under-resourced and rural medical environments.
Independent medical experts reviewing the study noted that the integration of machine learning into wound care analytics represents a crucial step forward. By recognizing that biological markers like albumin and physical characteristics like surface area do not operate in isolation, clinicians can eventually transition from reactive treatments to predictive, individualized care plans.
Future Outlook & Clinical Implications
As healthcare systems grapple with the mounting financial and logistical pressures of chronic wound management—particularly in aging populations and rural areas with limited healthcare access—the implications of this study are far-reaching.
Future advancements in wound care prognostication will rely heavily on:
- Multicenter Prospective Cohorts: Moving beyond single-center retrospective limitations to validate predictive algorithms across diverse geographic and socio-economic demographics.
- Time-to-Event Analytics: Incorporating longitudinal survival and recurrence models (such as Cox proportional hazards regression) to track wound trajectory over extended periods exceeding 12 weeks.
- Algorithmic Clinical Integration: Embedding machine learning classifiers like the validated MLP model into electronic health record (EHR) systems. This will allow clinicians to input real-time biochemical markers (like albumin) and physical measurements upon intake, instantly generating personalized risk profiles.
Ultimately, by bridging the gap between biochemistry, clinical interventions, and modern artificial intelligence, researchers are laying the groundwork for a new era of precision wound care—one that promises to alleviate patient suffering and optimize the allocation of scarce medical resources.
